Elastic Grid Reservations with User-Defined Optimization Policies

We present an algorithm for reserving compute resources in a Grid, that allows users to define an optimization strategy, if multiple candidates match the specified requirements. Reservations need not to be exact, but can be elastic in some properties like start-, end-time, duration or the number of requested CPUs. Our algorithm takes the speedup of applications into account to accordingly adjust the duration of the reservation when changing the number of CPUs. This new reservation approach allows users to more accurately specify their requirements and gives the computer system more flexibility to find appropriate resources than existing solutions. Caused by the enhanced flexibility, a single request can have many solutions to select from. To perform this selection efficiently in both cases, successful and failed requests, was one of our goals and we show performance results in Section 4. This work can also build as a basic block for flexible co-reservations – the allocation of multiple independent resources. Due to the elasticity of our reservation requests, the probability to find a common time-slot on all requested resources is higher with elastic requests than with fixed requests.